A machine learning-driven approach to predicting thermo-elasto-hydrodynamic lubrication in journal bearings

• Published in: Tribology international Vol. 196; p. 109670
• Main Authors: Cartwright, Samuel, Rothwell, Benjamin C., Figueredo, Grazziela, Medina, Humberto, Eastwick, Carol, Layton, James, Ambrose, Stephen
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01-08-2024
• Subjects: Hydrodynamic lubrication; Journal bearing; Machine learning; Neural network; OpenFOAM; PyTorch; OpenFOAM; Hydrodynamic lubrication; PyTorch; Neural network; Machine learning; Journal bearing
• ISSN: 0301-679X; 1879-2464
• DOI: 10.1016/j.triboint.2024.109670

Traditional methods of evaluating the performance of journal bearings, for example thermal-elastic-hydrodynamic- lubrication theory, are limited to simplified conditions that often fail to accurately model real-world components. Numerical models that include additional phenomena such as cavitation and fully coupled effects like deformation, temperature, pressure and viscosity can be more accurate but require a large amount of computational overhead, making analysis slower and more costly. To address this limitation, a novel machine learning-driven approach is developed to predict the 2D distribution of surface deformation, film thickness, temperature, and pressure across the bearing surface as a function of design variables such as load and speed. The training dataset, generated using a fully coupled Reynolds’ Equation solver implemented in OpenFOAM, contains a significantly extended range of conditions than in previous studies with approximately 39000000 points encompassing 4925 different test cases. Modelled bearing speeds range from 2000 to 10000rpm, while load values are varied between 1 and 30kN. Predicting surface deformation, film thickness, temperature and pressure across the bearing surface results in a mean absolute percentage error below 0.4% or better. The work also demonstrates that the trained models have a strong ability to generalise the prediction beyond the original training data range with only a 1% error at up to 200% of the highest trained speed. This work also demonstrates that machine learning-based processes are a practical alternative to physics-based numerical modelling, especially in cases where rapid performance evaluation is desired as real-time calculation is possible with significantly reduced computational cost. This has the potential to enable development of rapid design optimisation tools and real-time performance monitoring at high resolution and with low latency. Using consumer hardware, it is found that the neural network-based approach is faster than the existing numerical modelling technique by a factor of over 10000, enabling real-time predictions of lubrication systems.
•A neural network approach is demonstrated, using independent models to predict each variable.•Neural network models trained using an OpenFOAM numerical model display a mean error of <0.4%.•Once trained, the neural network models are ∼10 000 x faster than traditional numerical models.